Appearance
How the AI Works
What actually happens when a message or comment arrives, what the AI is given, what it keeps, and where it deliberately stays out of the way.
This is the conceptual companion to the AI Features guide — read that one first if you just want to switch things on.
The one-sentence version
Every incoming DM and comment is shown to a language model, which answers one narrow question about it, and InfluDM acts on that answer.
Two features, two questions:
| Feature | The question | The action |
|---|---|---|
| Collab Detection | Is this a brand proposing a collaboration? | Save it to Brand Collabs |
| Comment Moderation | Which of five kinds is this comment? | Delete it, or log it |
Nothing else is asked and nothing else is inferred. The AI doesn't write your replies, doesn't decide who to message, and never sends anything on your behalf.
The order things happen in
The sequence matters, because it's what makes the cost predictable:
- A message or comment arrives from Instagram.
- The flow engine sees it first. If one of your flows is already handling this message — an answer to a lead form, a tap on your opening message — it's consumed there and the AI never sees it.
- The gates are checked. Is the feature on? Is this Instagram account in its list? Is there enough balance? Any "no" stops here, and nothing is charged.
- The AI reads it and answers its one question.
- Credits are deducted, once, for that read.
- The action runs — a row is saved, or a comment is deleted.
Step 3 is why a workspace with the features switched off costs nothing at all, and step 5 is why you're never charged twice for the same message.
Reading is charged, acting is free
You're charged for the reading, not the outcome. A DM the AI reads and correctly decides is not a collab request costs the same as one it saves. That's the honest way round — the work was done either way, and a system that only charged for "hits" would have an incentive to find them.
Why it isn't keyword matching
The obvious way to build both features is a word list. It doesn't survive contact with a real Instagram account.
For collabs, the words that appear in a genuine offer — collab, partnership, brand — appear just as often in a pitch from someone who wants you to promote them. The distinction is direction: who is offering whom what. That isn't visible in a word list.
For comments, a word list is worse. Instagram slang inverts constantly — sick, insane, this killed me, stop it are praise. A filter matching those deletes your most enthusiastic fans. Meanwhile actual abuse routes around any list within a day, using spacing, emoji and transliteration.
Reading meaning handles both, and keeps handling them as language moves.
What the AI is given
Only the text of the single message or comment being classified, plus the instruction describing the question.
It is not given your other messages, your contact list, your leads, your flows, your account credentials or anything about your business. Each read is independent and stateless — the AI has no memory of the previous message and no way to reach anything else in your workspace.
What gets stored
| Feature | Stored | Not stored |
|---|---|---|
| Collab Detection | Messages identified as collab offers, with the sender and the offer | Every other DM. Read, classified, discarded. |
| Comment Moderation | Comments classified as Spam, Slurs/Abuse, Negative or Criticism | Neutral comments — the large majority — are never written down at all. |
The moderation log keeps roughly a month of history and then clears itself.
Dropping Neutral is deliberate and does two useful things at once: it keeps the log readable, and it means the overwhelming bulk of your comment traffic leaves no trace in our database. See Data & Privacy.
Where the AI can be wrong
It's a judgement, not a lookup, so treat it like one.
On collabs, the failure mode is a missed offer rather than a false one — a message that mentions a brand in passing is usually read as conversation. If you think something was missed, the DM is still in your Instagram inbox; nothing is hidden or moved.
On comments, the failure mode matters more, because deletion is permanent. This is exactly why the default delete list is narrow, why Criticism is off by default, and why everything the AI would have deleted under a wider setting sits in the log as Logged only. The log is the safety net — read it for a week before widening what you delete.
Offer amounts are read from ordinary human sentences, which are ambiguous more often than you'd think. "20k for 3 reels" is a total to some brands and a rate to others. The amount is editable for this reason, and the brand's original wording is always kept alongside it.
What it costs
A credit or two per message or comment, depending on length. The balance is prepaid and can never go negative, so the worst case is that reading stops — your flows are never affected.
AI Billing & Usage has the full picture.
Related
- AI Features — the practical guide.
- How a Flow Runs — the automation engine the AI sits beside.
- Data & Privacy — what's stored, for how long.
- Glossary — terms used across these docs.